JP2020501238A - Face detection training method, apparatus and electronic equipment - Google Patents

Abstract

Embodiments of the present invention provide a face detection training method, apparatus, and electronics that obtain a batch data training sample of the current iteration and determine a central loss value corresponding to each training sample. Determining, based on the central loss value corresponding to each training sample, determining the central loss value corresponding to the batch data training sample; and at least based on the central loss value corresponding to the batch data training sample. To determine the target loss value and, if the target loss value for face detection does not reach the convergence condition, update the network parameters in the face detection model based on the target loss value for face detection and proceed to the next iteration And outputting face detection when the target loss value for face detection reaches the convergence condition. According to an embodiment of the present invention, it is possible to enable face detection to guarantee high inter-class detection performance for faces and non-faces, while having invariance to differences within classes of faces, Robustness can be improved.

Description

  This application was filed with the China Patent Office on June 2, 2017, and filed with the application number "201710406726.9" and the title of the invention "Face Detector Training Method, Apparatus and Electronic Equipment". It claims priority from a patent application, the entire contents of which are incorporated by reference into this application.

  The present invention relates to the field of image processing, and more particularly, to a face detection training method, apparatus, and electronic apparatus.

  Face detection is a technology for detecting a face from an image using a face detector. Whether face detection training is good or bad directly affects the face detection effect. Optimization has always been the focus of research by those skilled in the art.

  With the development of deep learning, face detection training based on CNN (Convolutional Neural Network, convolutional neural network), for example, using a convolutional neural network that uses face region training based on Faster RCNN (Faster Region-based Convolutional Neural Network). Performing face detection is a mainstream training method for face detection. The face detection training process based on CNN mainly constructs a face detection model, performs repetitive training using training samples, and performs face detection model Update the network parameters to achieve face detection training optimization, and note that at each iteration the network parameters of the face detection model Can be regarded as face detection optimization processing.

  The current goal of face detection optimization is to maximize the difference between faces and non-faces (ie, to maximize the differences between classes), and the differences between faces Since attention is not paid, face detection has low discrimination ability and poor robustness when dealing with face changes in different scenes.

  In view of the above, the embodiments of the present invention provide a face detection training method, apparatus, and electronic device for improving face detection discrimination ability of face detection and improving robustness of face detection.

  To achieve the above object, embodiments of the present invention provide the following technical solutions.

A face detection training method,
Obtaining a batch data training sample of the iteration, wherein the batch data training sample includes a plurality of training samples of different sample classes;
Determining a central loss value corresponding to each training sample based on a feature vector of each training sample and a central feature vector of a sample class to which each training sample belongs;
Determining a central loss value corresponding to the batch data training sample based on the central loss value corresponding to each of the training samples;
Determining a target loss value for face detection based on the central loss value corresponding to the batch data training sample;
And outputting a face detection training result when the target loss value for face detection reaches a set training convergence condition.

Embodiments of the present invention further provide a face detection training device,
A sample acquisition module for acquiring a batch data training sample of this iteration, wherein the batch data training sample includes a plurality of training samples of different sample classes;
A sample center loss value determination module for determining a center loss value corresponding to each training sample based on a feature vector of each training sample and a center feature vector of a sample class to which each training sample belongs;
A batch sample center loss value determination module for determining a center loss value corresponding to the batch data training sample based on the center loss value corresponding to each of the training samples;
A detection target loss value determination module for determining a target loss value for face detection based on at least the central loss value corresponding to the batch data training sample;
If the target loss value of the face detection does not reach the set training convergence condition, the network parameter of the face detection model is updated based on the target loss value of the face detection, and the parameter is updated to proceed to the next iteration. Modules and
A detection output module for outputting a face detection training result when the target loss value for face detection reaches a set training convergence condition.

An embodiment of the present invention further provides an electronic device, including a memory and a processor,
A program is stored in the memory, and the processor calls the program, and according to the program,
Obtaining a batch data training sample of this iteration, wherein said batch data training sample includes a plurality of training samples of different sample classes;
Determining a central loss value corresponding to each training sample based on a feature vector of each training sample and a central feature vector of a sample class to which each training sample belongs;
Determining a central loss value corresponding to the batch data training sample based on the central loss value corresponding to each of the training samples;
Determining a target loss value for face detection based on at least the central loss value corresponding to the batch data training sample;
If the target loss value of the face detection has not reached the set training convergence condition, based on the target loss value of the face detection, update the network parameters of the face detection model, proceed to the next iteration,
When the target loss value of the face detection reaches the set training convergence condition, the detection result of the face detection is output.

  Embodiments of the present invention further provide a computer readable storage medium including instructions, and causing the computer to perform the method described in the first aspect when the instructions are executed on the computer.

  Embodiments of the present invention further provide a computer program product that includes instructions, which when executed on a computer, causes the computer to perform the method described in the first aspect.

  According to the above technical solution, the face detection training procedure provided by the embodiment of the present invention can include: obtaining a batch data training sample of the current iteration; The sample includes a plurality of training samples of different sample classes, and determines a central loss value corresponding to each training sample based on a feature vector of each training sample and a central feature vector of a sample class to which each training sample belongs. A central loss value corresponding to the batch data training sample is determined based on the central loss value corresponding to each of the training samples, and a target loss value for face detection is determined based on at least the central loss value corresponding to the batch data training sample. Confirmation, said face inspection If the target loss value does not reach the set training convergence condition, the network parameters in the face detection model are based on the target loss value for face detection until the target loss value for face detection reaches the set training convergence condition. Is updated to the next iteration, and when the target loss value of the face detection reaches the set training convergence condition, the face detection is output, and the face detection training is completed.

  In an embodiment of the present invention, by combining the face optimization training goal with a central loss value corresponding to a batch data training sample, face detection is insensitive to intra-class differences between faces. Since it is possible to have degeneration, by performing the face detection optimization training by combining the central loss values corresponding to the batch data training samples, the optimization-trained face detection is a Since it has invariance, it is possible to guarantee high inter-class detection performance for faces and non-faces, and it is possible to improve the robustness of face detection.

  In order to more clearly describe the technical solutions in the embodiments of the present invention or the related art, the drawings used in the description of the embodiments or the related art will be briefly described below. Apparently, the drawings in the following description are merely embodiments of the present invention. For those skilled in the art, other drawings may also be derived from the drawings provided without creative effort.

It is a structure of a face detection model. This is another configuration of the face detection model. FIG. 3 is a block diagram of a hardware configuration of the electronic device. 5 is a flowchart of a face detection training method provided by an embodiment of the present invention. It is a schematic diagram of face detection training based on a face detection model. It is a flowchart of the determination method of a face frame coordinate regression loss value. It is a flowchart of the acquisition method of a batch data training sample. 1 is a block diagram illustrating a configuration of a face detection training apparatus provided by an embodiment of the present invention. FIG. 3 is a block diagram illustrating another configuration of the face detection training apparatus provided by an embodiment of the present invention. FIG. 4 is a block diagram illustrating another configuration of the face detection training apparatus provided by an embodiment of the present invention.

  The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are merely some but not all of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments in the present invention without creative efforts fall within the protection scope of the present invention.

  The embodiment of the present invention is a selectable face detection model constructed based on CNN, and includes a basic network layer, a candidate frame prediction layer, and a face detection layer as shown in FIG.

  The basic network layer may be a sub-network that is sequentially connected by a series of convolution layers and a pooling layer, and the basic network layer may be formed by a series of convolution layers so that each of the training samples ( The training sample may be a sample in the form of an image), and a convolution process for each layer can be performed on the convolution process. In the next convolution layer, the convolution process result output from the previous convolution layer is convolved. Here, in the image features processed by the multiple convolutional layers, the features of the shallow layer can represent features such as rich edge corners, texture structures, and the features of the deep layer can be further based on the features of the shallow layer. This is an abstract mapping, and the convolution process for each layer of the multi-layer convolution layer Therefore, extraction of image features of different layers can be realized, and for each training sample, the output of the basic network layer is a feature map convolved by the convolution layer of the last layer. , The feature map is an expression of an image feature.

  The candidate frame prediction layer may be a full convolution subnetwork constructed on the image features output by the base network layer, wherein the candidate frame prediction layer maps the features of each training sample by the convolution layer. , Thereby constructing a candidate frame classifier and a candidate frame regressor with the mapped nodes, forming a candidate frame detection, and using the candidate frame classifier to make a probability prediction of the candidate frame (Proposals). A candidate frame regressor can be used to predict the coordinates of a candidate frame, thereby outputting candidate frames (Proposals) and inputting the candidate frame output from the candidate frame prediction layer to the face detection layer. can do.

  The face detection layer may be a subnetwork including an attention area pooling layer (RoI Pooling) constructed based on the image features output from the basic network layer and the candidate frames output from the candidate frame prediction layer. On the other hand, the face detection layer performs dimension reduction sampling on the image feature of the training sample output from the basic network layer based on the candidate frame (Proposals), acquires a fixed-size feature map, All nodes in the map can be mapped to a fixed-length feature vector, which gives a feature vector for each training sample and builds a face classifier and face regressor based on the feature vector for each training sample Then, the face classifier and the face regressor combine to form face detection. , The face classifier can predict face and non-face probabilities, and the face regressor can perform more accurate coordinate regression of the face frame based on the candidate frame.

  In addition, a further selectable subdivision of the face detection model shown in FIG. 1 can be realized by a face detection model based on the Faster RCNN, which is a typical algorithm for face detection and the RPN (Region) It is divided into a Proposal Networks layer and a Fast RCNN layer. The RPN layer generates a candidate frame, and the Fast RCNN layer can obtain final test results based on the candidate frame.

  As shown in FIG. 2, a face detection model based on the Faster RCNN can include a basic network layer, an RPN layer, and a Fast RCNN layer, in which the RPN layer is considered to be a selectable realization of a candidate frame prediction layer. The Fast RCNN layer can be considered as a selectable realization of the face detection layer.

  In the embodiment of the present invention, the goal of the RPN layer is to generate a candidate frame based on the image features output by the basic network layer. In this process, the embodiment of the present invention uses a plurality of anchor frames in advance. The plurality of anchor frames can cover different scales and aspect ratios, the sub-frames in the training sample can be determined by the predefined anchor frames, and the candidate frames can be defined by the sub-frames. It can be predicted (for example, by training the candidate frame detection using the sub-frame, the candidate frame can be predicted by the candidate frame detection).

  Also, the anchor frame is inside the RPN layer and may be used to define and construct a classifier and regressor for the Proposal. RPN is candidate frame detection. Specifically, each anchor frame is associated with a respective detection (classification and regression), and classification and regression require predicted and target values for training and learning. In RPN, the determination of the classification target value (ie, how to define whether this output is a positive class or a negative class) is based on the overlap rate between the anchor frame and the real frame. Similarly, in the Fast RCNN, the determination of the classification target value is based on the overlap rate between the candidate frame and the actual frame. Therefore, the anchor frame used by the RPN and the candidate frame used by the Fast RCNN have a similar effect when constructing a classifier, and the anchor frame can be regarded as a candidate frame of the candidate frame. The RPN can build multiple candidate frame detections for each node after image feature convolution has been performed (each candidate frame detection is associated with one anchor frame).

  The goal of the Fast RCNN layer is to generate a feature vector of the training sample based on the candidate frame and the image features output by the basic network layer, thereby using a face classifier and a face regressor with the feature vector of the training sample. And the face classifier and the face regressor combine to form face detection.

  In order to have a better detection effect on face detection, iterative training can be performed by a model optimization algorithm such as a stochastic gradient descent algorithm (SGD), and in each iteration, batch data training is performed from a training sample set. Training is performed by selecting samples, and then, in each iteration, the network parameters of the face detection model are updated depending on whether the face detection optimization goal has been achieved.

  Currently, the main goal of face detection optimization is to maximize the difference between a face and a non-face, ignoring differences in face changes in different scenes between the face and the face, for example, Differences in face changes in scenes such as shooting angles, resolutions, lighting conditions, changes in facial expressions, and occlusion are ignored, and the discrimination ability of trained face detection is weakened, resulting in poor robustness. For example, if the difference between a face and a face (for example, with or without light) in a class is too large, it is determined as a different class, but it is actually the same class. , The differences in the classes need to be as small as possible to ensure that face detection is invariant to the differences in the classes.

  Based on this, embodiments of the present invention improve the iterative training optimization process of face detection and submit a new face detection training method whereby face detection is performed within the class between faces. While reducing the difference, it is ensured that the face and the non-face have high detection performance, and the face detection discrimination ability is improved.

  The face detection training method provided by an embodiment of the present invention can be loaded into an electronic device for performing face detection training in the form of a program, and the electronic device can be a server on a network side. Alternatively, the terminal device may be a terminal device such as a personal computer (PC) on the user side, and the form of the electronic device can be determined according to actual training needs for face detection.

  As shown in FIG. 3, the hardware configuration of the electronic device for performing face detection training includes at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4. Can be included.

  In the embodiment of the present invention, the number of the processor 1, the communication interface 2, the memory 3, and the communication bus 4 is at least one, and the processor 1, the communication interface 2, and the memory 3 communicate with each other via the communication bus 4. Obviously, the communication connections of processor 1, communication interface 2, memory 3, and communication bus 4 shown in FIG. 3 are merely optional.

Further, the communication interface 2 may be a communication module interface such as a GSM (registered trademark) module interface,
The processor 1 may be a central processing unit CPU, or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement an embodiment of the present invention. .

  The memory 3 may include a high-speed RAM memory, and may also include, for example, a non-volatile memory (NVM) such as at least one magnetic disk memory.

  Note that a program is stored in the memory 3, and the processor 1 calls up the program stored in the memory 3, and the program executes the face detection training method provided by the embodiment of the present invention.

  Embodiments of the present invention can perform iterative face detection training using a model optimization algorithm such as a stochastic gradient descent algorithm (SGD), and SGD is a commonly used convolutional neural network optimization algorithm. Yes, and effective in solving large-scale machine learning problems, SGD uses gradient descent optimization using batch data training samples (Minibatch) randomly extracted from a training sample set at each iteration. I do.

  FIG. 4 shows a flow of the face detection training method provided by the embodiment of the present invention, taking the face detection training according to one iteration as an example, and the flow of the face detection training according to each iteration is shown in FIG. Can be referred to. Referring to FIG. 4, the face detection training method provided by an embodiment of the present invention may include:

  Step S100, acquiring a batch data training sample of the current iteration, wherein the batch data training sample includes a plurality of training samples of different sample classes.

  In addition, the batch data training sample (Minibatch) may be extracted from a training sample set including all the training samples.

  Face detection can be thought of as fulfilling two categories of tasks (face and non-face). In each iteration, multiple face images are acquired from the training sample set as positive class training samples, and multiple non-face The image is acquired as a training sample of a negative class, and the acquired training sample of the positive class and the training sample of the negative class can constitute a batch data training sample for each iteration.

  Correspondingly, the batch data training sample used in the current iteration may include a plurality of training samples, and the sample class of the plurality of training samples is a positive class (ie, a training class in which a face image is a positive class). Sample) and a negative class (that is, corresponding to a training sample having a non-face image as a negative class).

  In step S110, a central loss value corresponding to each training sample is determined based on a feature vector of each training sample and a central feature vector of a sample class to which each training sample belongs.

  For one training sample in the batch data training sample, an embodiment of the present invention determines a feature vector of the training sample and a corresponding central feature vector in the batch data training sample of a sample class to which the training sample belongs. Whereby a central loss value corresponding to the training sample can be determined, the process being performed for each training sample in the batch data training sample, and for each training sample in the batch data training sample. To obtain the central loss value corresponding to.

  Also, the central feature vector of one sample class in the batch data training sample may be updated in this iteration corresponding to the average value of the feature vectors of the training samples belonging to the sample class among the batch data training samples. Good.

  In addition, for one sample class, the embodiment of the present invention determines each training sample belonging to the sample class in the batch data training sample, and determines a characteristic of each training sample belonging to the sample class in the batch data training sample. Based on the vector, determine the average value of the feature vector of each training sample belonging to the sample class, obtain an update variable of the central feature vector of the sample class of the batch data training samples, and set the update variable. And obtaining a central feature vector of the sample class in the batch data training sample based on the learning rate. Based on the average value of the feature vectors of training samples, it may be implemented to update the central feature vector of the sample class.

  Further, the embodiment of the present invention can determine the center feature vector of one sample class in the batch data training sample by the following equation.

Among them,
Represents the set learning rate,
Represents an update variable,
Represents the central feature vector corresponding to the batch data training sample used by the jth sample class at the tth iteration,
Represents the central feature vector corresponding to the batch data training sample used by the j-th sample class in the t + 1 iteration.

  That is, for one sample class, the embodiment of the present invention subtracts the product of the update variable and the set learning rate from the central feature vector corresponding to the sample class of the batch data training sample in the previous iteration. Then, the central feature vector corresponding to the sample class of the batch data training sample in this iteration may be obtained.

  Further, with respect to the sample class of the positive class in the batch data training sample, the embodiment of the present invention determines the feature vector of each training sample belonging to the positive class among the batch data training samples, and determines each training sample belonging to the positive class. By determining the average value of the feature vectors of the positive class, the central feature vector of the sample class of the positive class can be updated, and correspondingly, for the sample class of the negative class in the batch data training sample, the present invention is implemented. In the example, the feature vector of each training sample belonging to the negative class among the batch data training samples is determined, and the feature vector of each training sample belonging to the negative class is determined. By determining the average value of Le, it is possible to update the central feature vectors of sample class negative class.

  Further, for each training sample of the positive class among the batch data training samples, the embodiment of the present invention, based on the feature vector of each training sample of the positive class and the center feature vector of the sample class of the positive class, A central loss value corresponding to each training sample of the positive class can be determined, and for each training sample of the negative class among the batch data training samples, an embodiment of the present invention provides a feature vector and a feature vector of each training sample of the negative class. , The central loss value corresponding to each training sample of the negative class can be determined based on the central feature vector of the sample class of the negative class.

The center loss value of one training sample, the feature vector of the training samples may be represented by the distance between the center feature vectors of sample class to which the training samples belong, x _i is out of the batch data training samples represents the _ith training sample, y _i represents the sample class to which x _i belongs (y _i = 1 represents the positive class, y _i = 0 can be set to represent the negative class, and of course, y _i = 0 represents a negative class, y _i = 1 can be set to represent a positive class, and y _i values corresponding to the positive class and the negative class may be different), and c _yi belongs to x _i assuming represents the center feature vectors of sample class y _i, heart loss of x _i samples The value
Can be defined as

  Note that, in the process of determining the feature vector of one training sample, after the basic network layer outputs the image features of the training sample, the face detection layer determines the region of interest in the training sample based on the candidate frame (Proposals), The face detection layer performs dimension reduction sampling on the image feature of the attention area of the training sample, acquires a fixed-size feature map, connects all the nodes in the feature map, and maps it to a fixed-length feature vector. , A feature vector of the training sample may be obtained.

  Step S120, determining a central loss value corresponding to the batch data training sample based on the central loss value corresponding to each training sample.

  Further, according to the embodiment of the present invention, based on the central loss value corresponding to each training sample, the average value of the central loss value corresponding to each training sample is determined, and the average value of the central loss value corresponding to each training sample is determined. Based on this, a central loss value corresponding to the batch data training sample may be determined.

  Further, in the embodiment of the present invention, the average value of the center loss values corresponding to each training sample may be used as the center loss value corresponding to the batch data training sample, and the center loss value corresponding to each training sample may be used. May be multiplied by a set value (eg, 例 え ば) to obtain a central loss value corresponding to the batch data training sample.

Assuming that the batch data training sample has m training samples, the central loss value corresponding to the batch data training sample is
Can be expressed as

  Step S130, determining a target loss value for face detection based on at least the central loss value corresponding to the batch data training sample.

  The target loss value for face detection is an expression of an optimization target in the face detection iterative training process. When the target loss value reaches a set training convergence condition (for example, a minimum), the iterative training ends and face detection is performed. And in each iteration, the embodiment of the present invention uses the existing face detection optimization goal as the face detection optimization goal of the embodiment of the invention to the batch data training sample to be used. A target loss value for face detection is obtained in combination with the corresponding center loss value.

  In addition, the embodiment of the present invention is based on the central loss value corresponding to the batch data training sample, the classification loss value corresponding to the batch data training sample, and the face frame coordinate regression loss value corresponding to the batch data training sample. Alternatively, the target loss value for face detection may be determined.

  The classification loss value corresponding to the batch data training sample can be determined based on the difference between the classification prediction probability and the classification target probability (true classification probability) of each training sample in the batch data training sample.

  In the embodiment of the present invention, for each training sample of the batch data training sample, after acquiring a feature vector of the training sample, a sample class to which the training sample belongs is predicted using a Softmax function or the like, and the A classification prediction probability can be obtained, and based on the classification prediction probability of the training sample and the true classification target probability of the training sample, the classification loss value corresponding to the training sample (for example, the classification prediction probability of the training sample and Difference from the classification target probability), and based on the classification loss value corresponding to each training sample in the batch data training sample, the classification loss value corresponding to the batch data training sample is determined. It is (for example, taking the average value of the classification loss value of each training sample).

  As can be seen, the classification loss value corresponding to the batch data training sample is an index by which face detection classifies between face and non-face classes, and the classification loss value corresponding to the batch data training sample is The difference between non-faces (difference between classes) can be represented, and the classification loss value corresponding to the batch data training sample is used as part of the optimization target for face detection, so that optimized face detection , High performance in discriminating between face and non-face classes.

  Based on this, the central loss value corresponding to the batch data training sample represents the distance between the feature vector of the training sample and the central feature vector of the sample class to which the training sample belongs, so that the central loss value corresponding to the batch data training sample The value can describe the difference between the feature vector of the training sample and the central feature vector of the sample class to which it belongs, and can represent the feature vector difference within the class of the training sample of each sample class, Using the central loss values corresponding to the batch data training samples as part of the optimization goal for face detection, the optimized face detection can be used to determine differences within a class of faces (eg, between faces in different scenes). Invariance within the class of It is possible to improve the robustness of the detection.

  In addition, face detection training may include classification training and regression training, and is a process of joint training, in which a single loss includes a central loss value and a classification loss value corresponding to a batch data training sample. The value can be regarded as an optimization target of the classification training, and as the optimization target of the classification training in the face detection training, for example, a loss value constituted by a central loss value sum and a class loss value corresponding to the batch data training sample. May be minimized.

  In each iteration, the optimization goal of regression training in face detection training can be configured by face frame coordinate regression loss values corresponding to batch data training samples.

  By combining the central loss value, the classification loss value, and the face frame coordinate regression loss value corresponding to the batch data training sample in one iteration, a target loss value for face detection is formed to optimize the face detection training target. Can be represented.

  The embodiment of the present invention further comprises a product of a center loss value corresponding to the batch data training sample and a first set weight, a face frame coordinate regression loss value corresponding to the batch data training sample and a second set weight. And the classification loss value corresponding to the batch data training sample to obtain a target loss value for face detection.

L _cls represents the classification loss value corresponding to the batch data training samples, the L _c is assumed to represent the center loss value corresponding to the batch data training samples, the target loss value of the face detection, L _{cls +} μL c ₊ λL _Expressed as _reg , μ and λ represent set weighting factors, where μ is a first set weight and λ is a second set weight.

  Further, in the embodiment of the present invention, the target loss value for face detection may be obtained by directly summing the central loss value, the classification loss value, and the face frame coordinate regression loss value corresponding to the batch data training sample. .

  Step S140: It is determined whether the target loss value of the face detection reaches the set training convergence condition. If not, step S150 is executed. If so, step S160 is executed.

  Further, the set training convergence condition can be regarded as minimizing the target loss value of face detection.

  Specifically, the smaller the classification loss value corresponding to the batch data training sample, the better the effect of face detection on the classification of face and non-face, and the face detection maximizes the distinction between face and non-face (ie, The difference between the feature vectors in the classes of the training samples in each sample class is smaller, the smaller the central loss value corresponding to the batch data training sample, the smaller the training sample of the same sample class. Can be reduced to further reduce the difference between faces in the sample class, i.e., by iterative training, the feature vector of each training sample in the batch data training sample and the sample class to which the training sample belongs Distance to the central feature vector of It is.

  As can be seen, the target loss value for face detection is determined by combining the central loss values corresponding to the batch data training samples, thereby determining the training convergence condition with the target loss value for face detection. Detection is based on the case where the face detection determines the difference in the class of the face (eg, the difference in the class between faces in different scenes) in the case of the central loss value corresponding to the minimized batch data training sample. On the other hand, it is possible to improve the robustness of face detection by guaranteeing invariance.

  Step S150: Update the network parameters related to face detection in the face detection model based on the target loss value of the face detection, proceed to the next iteration, and return to step S100.

  Further, if the target loss value for face detection does not reach the set training convergence condition (for example, the target loss value for face detection does not reach the minimum), the embodiment of the present invention will be described. Based on the values, the network parameters in the face detection model can be updated, and the next iteration is performed according to the flow of the iterative training, the process returns to step S100, and the result of the determination in step S140, the target loss value for face detection is set. Until the training convergence condition is reached, steps S100 to S140 may be repeatedly executed using the face detection model whose network parameters have been updated.

  Further, the embodiment of the present invention may proceed to the next iteration by the stochastic gradient descent method and return to step S100.

  Step S160: Output face detection.

  When the target loss value for face detection reaches a set training convergence condition (for example, when the target loss value for face detection is minimized), the face detection obtained by face detection model training can be output. Alternatively, the iterative training optimization process for face detection may be completed.

  The face detection training flow provided by an embodiment of the present invention may include: obtaining a batch data training sample of the current iteration, wherein the batch data training sample includes a plurality of different sample classes. And a center loss value corresponding to each training sample is determined based on a feature vector of each training sample and a center feature vector of a sample class to which each training sample belongs, and a center corresponding to each training sample is determined. Determining a central loss value corresponding to the batch data training sample based on the loss value; determining a target loss value for face detection based on at least the central loss value corresponding to the batch data training sample; Loss value set If the training convergence condition has not been reached, the network parameters in the face detection model are updated based on the target loss value for face detection until the target loss value for face detection reaches the set training convergence condition. When the target loss value for face detection reaches the set training convergence condition, face detection can be output, and training for face detection is completed.

  In an embodiment of the present invention, the face detection training optimization goal is combined with a central loss value corresponding to the batch data training sample, so that face detection is performed with respect to intra-class differences between faces. By performing face-training optimization training by combining the central loss values corresponding to the batch data training samples, it is possible to have invariance, so that the optimized-trained face detection reduces the differences within the class of faces. , It is possible to guarantee high inter-class detection performance for faces and non-faces, and to improve the robustness of face detection.

  Further, the embodiment of the present invention, when updating the network parameters in the face detection model based on the target loss value of the face detection, based on the target loss value of the face detection, by back propagation, the network parameters in the face detection model May be updated.

  Further, the embodiment of the present invention determines the parameter update value of the face detection based on the target loss value of the face detection and the network parameter in the face detection model of the previous iteration, thereby obtaining the parameter update value of the face detection. , The network parameters in the face detection model of the previous iteration may be updated.

Assuming that the target loss value of face detection is Loss, Loss = L _ls + μL _c + λL _reg , and the network parameter in the previously repeated face detection model is W1, the parameter update of face detection is performed. value is,
Can be expressed as

Updating the network parameters in the previously repeated face detection model based on the face detection parameter update value can be realized by the following equation.

W2 is a network parameter of the updated face detection model, k is a moving amount,
Is a learning rate, and s is a weight attenuation coefficient.

  In addition, according to the embodiment of the present invention, as shown in FIG. 5, a center loss function (Center Loss) can be installed in a face detection layer (for example, a Fast RCNN layer), and the center loss function is determined by the face detection layer. Of the fully connected feature representation layer, which connects all nodes in the feature map in a fully connected manner to obtain a feature vector of each training sample, and has a fixed length , So that in each iteration training, the central loss function is based on the feature vector of each training sample of the batch data training samples used in this iteration, and for each of the batch data training samples Determine the central loss value corresponding to the training sample and set the batch data training sample The central loss value Lc corresponding to may be determined correspondingly.

At the same time, a Softmax function can be installed in the face detection layer (for example, the Fast RCNN layer), and the Softmax function can be applied to the fully connected feature representation layer of the face detection layer. The function can process the feature vector of each training sample to determine the classification prediction probability of each training sample, and furthermore, classify the training sample classification prediction probability and classification target probability (classification target probability) by Softmax Loss (classification loss function). (True probability) and determine a classification loss value L _cls corresponding to the batch data training sample.

  That is, the input of the Softmax function is the feature vector of the training sample, the output is the prediction probability belonging to each sample class of the training sample, and the input of Softmax Loss (classification loss function) is p (classification prediction probability) of the training sample. The output is a loss value (Loss), and the smaller the Loss, the more accurate the classification. In an embodiment of the present invention, the Center Loss and the Softmax Loss operate on the same layer (that is, the input feature vectors are the same), the Center Loss is used as an auxiliary monitoring signal for optimizing face detection, and the Center Loss is used as an auxiliary monitoring signal. The smaller the value, the smaller the difference between the features in the classes detected by the face detection. The Softmax Loss separates the features between the classes detected by the face detection from each other, and guarantees a identifiable difference between the classes.

  Further, in the embodiment of the present invention, a face frame regression prediction function SmoothL1 (smoothed norm function) can be set in a face detection layer (for example, a Fast RCNN layer), and based on the candidate frame, A face frame predicted coordinate corresponding to each training sample in the batch data training sample is determined, and further, a face frame coordinate regression loss value corresponding to each training sample is determined by SmoothL1 Loss, and the input is a face frame corresponding to the training sample. The predicted coordinates and the face frame target coordinates, and the output is a loss value (Loss). Next, the face frame coordinate regression loss value Lreg corresponding to the batch data training sample is determined.

Further, in the embodiment of the present invention, the target loss value Loss = L _ls + μL _c + λL _reg for face detection is determined, and the target loss value Loss obtained in each iteration until the target loss value Loss is minimized. , The network parameters in the face detection model may be updated.

  The process of determining the classification loss value corresponding to the batch data training sample in one iteration is as follows.

Determining a classification loss value corresponding to each training sample in the batch data training sample, based on the classification prediction probability and the classification target probability corresponding to each training sample in the batch data training sample;
A classification loss value corresponding to the batch data training sample may be determined based on a classification loss value corresponding to each training sample in the batch data training sample.

  In addition, the determination processing of the face frame coordinate regression loss value corresponding to the batch data training sample in one iteration can include the following as shown in FIG.

  In step S200, based on the candidate frame, the face frame predicted coordinates corresponding to each training sample in the batch data training sample are determined.

  Further, the embodiment of the present invention determines the attention area of each training sample of the batch data training sample in the current iteration based on the candidate frame output from the candidate frame prediction layer, and determines the face frame prediction corresponding to each training sample. The coordinates can be obtained, and the predicted coordinates of the face frame of the training sample can be represented by the abscissa of the upper left vertex, the ordinate of the upper left vertex, the abscissa of the lower right vertex, the ordinate of the lower right vertex, and the like.

  Further, in the embodiment of the present invention, a face frame regression prediction function SmoothL1 (smoothing norm function) is installed on a face detection layer (for example, a Fast RCNN layer), and each training sample is supported by SmoothL1 based on a candidate frame. May be determined.

  Step S210, a face frame coordinate regression loss value corresponding to each training sample is determined based on the face frame predicted coordinates corresponding to each training sample and the face frame target coordinates corresponding to each training sample.

  Further, the face frame target coordinates corresponding to the training sample may be coordinates that truly correspond to the face frame in the training sample, and for each training sample, the embodiment of the present invention may use the face frame corresponding to the training sample. The regression loss value of the face frame coordinates corresponding to the training sample can be determined from the difference between the predicted coordinates and the target coordinates of the face frame. The face frame coordinate regression loss value can be obtained.

  Also, in the embodiment of the present invention, the face frame coordinate regression loss value can be represented by SmoothL1 Loss, the input is the face frame predicted coordinate and the face frame target coordinate corresponding to the training sample, and the output is the loss value (Loss). ), Indicating that the smaller the Loss is, the more accurate the regression of the face frame is.

  Step S220, determining a face frame coordinate regression loss value corresponding to the batch data training sample based on the face frame coordinate regression loss value corresponding to each training sample.

  Further, the embodiment of the present invention determines the average value of the face frame coordinate regression loss values corresponding to each training sample based on the face frame coordinate regression loss values corresponding to each training sample in the batch data training sample, and determines the average value. Based on the value, a face frame coordinate regression loss value (SmoothL1 Loss) corresponding to the batch data training sample may be determined.

  Further, in the embodiment of the present invention, the process of performing face detection iterative training uses a multi-loss function joint training including two joint tasks of face classification and regression, and the classification training uses Center Loss and Softmax Loss. And regression training are optimized using SmoothL1 Loss, and the final optimization goal for face detection is the three loss values of Center Loss, Softmax Loss, and SmoothL1 Loss corresponding to the batch data training samples. Is to be minimized.

  Embodiments of the present invention also provide for fine-tuning (Finetuning) pre-trained models in common large-scale face identification tasks (ImageNet), and introducing center loss values as auxiliary optimization goals for face detection. Guidance in optimizing and training face detection models may be used to improve face detection discrimination capabilities for intra-class differences between faces.

  In the iterative training process, the embodiment of the present invention determines a training sample that is hard to be detected by face detection in a training sample set based on the face detection model of the previous iteration, and determines a batch used in the next iteration. Determine the data training samples to improve the detection capability of face detection for these hard-to-detect training samples, and determine if the training samples are difficult to detect by measuring the target loss value corresponding to the training sample Explain that the higher the target loss value, the farther the training sample is from the optimization target and the more difficult the detection is.

  Correspondingly, FIG. 7 shows a flow chart of a method for obtaining a batch data training sample of the current iteration provided by an embodiment of the present invention, and referring to FIG. 7, the method comprises: Can be included.

  Step S300, fixing the face detection model of the previous iteration, and acquiring the center loss value, the classification loss value, and the face frame coordinate regression loss value corresponding to each training sample in the training sample set by the face detection model of the previous iteration. I do.

  Step S310, determining a target loss value of each training sample in the training sample set based on the central loss value, the classification loss value, and the face frame coordinate regression loss value corresponding to each training sample in the training sample set.

  In addition, for one training sample, the embodiment of the present invention obtains a target loss value of the training sample by adding up the center loss value, the classification loss value, and the face frame coordinate regression loss value of the training sample. By performing this process on each training sample, the target loss value of each training sample may be obtained.

  Further, for one training sample, the target loss value may be represented as a classification loss value + μ center loss value + λ face frame coordinate regression loss value.

  Also, for one training sample, the embodiment of the present invention obtains the target loss value of the training sample by summing the center loss value, the classification loss value, and the face frame coordinate regression loss value of the training sample. Is also good.

  Step S320, selecting the first number of training samples having the largest target loss value among the sample classes of the positive class based on the target loss value of each training sample among the sample classes of the positive class in the training sample set; Based on the target loss value of each training sample among the sample classes of the negative class in the set, select a second number of training samples having the largest target loss value in the sample class of the negative class, and select the second number of training samples for the second number. The ratio value of the number 1 corresponds to the set proportion.

  In addition, the embodiment of the present invention can classify each training sample in the training sample set according to the sample class of the positive class and the negative class after obtaining the target loss value of each training sample in the training sample set. The target loss value of each training sample among the sample classes belonging to the positive class in the training sample set, and the target loss value of each training sample among the sample classes belonging to the negative class in the training sample set can be determined, The training samples belonging to the positive class are sorted based on the target loss value of each training sample among the sample classes of the positive class (the target loss values are sorted in ascending order). May be, it may be sorted target loss value in ascending order), based on the target loss value for each training sample in the sample class negative class may sort the training samples belonging to the negative class.

  Further, according to the set proportion between the training sample of the positive class and the training sample of the negative class in the batch data training sample, the sample class of the positive class is determined based on the target loss value of each training sample among the sample classes of the positive class in the training sample set. And selecting the first number of training samples with the largest target loss value among the sample classes of the negative class based on the target loss value of each training sample among the sample classes of the negative class in the training sample set. Selecting a training sample of a second number of loss values, wherein a ratio of the first number to the second number is required in the batch data training sample; So as to correspond to the number of set proportion of moths Restorative class sample.

  Also, based on consideration of the data balance requirement of the positive sample (face) and the negative sample (non-face) by Center Loss, the embodiment of the present invention can set the set proportion to 1: 1. The number 1 and the second number are the same.

  Step S330, the training sample selected from the sample class of the positive class and the training sample selected from the sample class of the negative class constitute a batch data training sample of the current iteration.

  As described above, in the face detection training method provided by the embodiment of the present invention, after sending the batch data training sample of the previous iteration to the face detection model for training, the Center Loss of the batch data training sample of the previous iteration and Based on Softmax Loss, the face detection is updated and optimized, and based on the SmoothL1 Loss of the batch data training sample of the previous iteration, the face regressor is updated and optimized, so that the face detection is performed in the Center Loss, Softmax Loss, Optimize for minimum weighted sum with SmoothL1 Loss.

  The previous iteration can determine the batch data training samples to be used in the next iteration, and the last iteration of the face detection model will determine the Center Loss, Softmax Loss, and SmoothL1 Loss of each training sample in the training sample set. A target loss value may be determined, whereby a training sample of a first number of positive classes with a maximum target loss value and a training sample of a second number of negative classes with a maximum target loss value are trained. Choosing from the set, the next iteration's Minibatch (ie, batch data training samples) is constructed.

  Accordingly, the process proceeds to the next iteration, and in the next iteration, the Minibatch is sent to the face detection model for training. In a certain iteration, the weighted sum of the smooth L1 Loss, the Softmax Loss, and the Smooth L1 Loss of the batch data training sample is minimized. Until then, repeat the training.

  In the training sample set described above, the training sample that is hardly detected by the face detection that was previously repeatedly trained is used as the Minibatch used in the next iteration, which makes it possible to better estimate the center loss in each iteration. Therefore, it is possible to better monitor and learn a feature having a discriminating power of a difference within a class in a training sample.

  What needs to be described here is that, unlike conventional iterative training, which uses a stochastic gradient descent algorithm to perform face detection, embodiments of the present invention simply use randomly sampled batch data training samples (Minibatch). Is not used to perform gradient descent optimization. In the previous iteration, a mini-batch to be used in the next iteration is determined by combining training samples that are hard to be detected from the training sample set.

  As can be seen, embodiments of the present invention provide a robust face detection training method. The method is realized by a neural network, and in each training process of the iterative training, a Center Loss (central loss value) corresponding to a batch data training sample is introduced as an auxiliary loss function of tasks of two categories, face and non-face. Then, the face detection optimization training can be monitored together with the Softmax Loss (classification loss value) corresponding to the batch data training sample, and the face detection learning process can be guided, whereby the face detection can be performed on the face and non-face. The present invention reduces differences within a class between a face and a face while maintaining the ability to distinguish the difference between the face and the class, and improves the discrimination ability of face detection for the face.

  Then, using the difficult sample online mining algorithm (OHEM), based on the total loss value of the training sample, the positive class training sample and the negative class training sample which are hard to detect in the previous training are mined, and the positive / negative sample is mined. Is maintained at 1: 1, which enhances the classification capabilities of face detection for training samples that are difficult to detect, and improves the overall performance of face detection.

  It should be noted that the present invention employs an anchor frame (covering multiple sizes and multiple aspect ratios) and a multi-scale training strategy that are more suitable for the facial target to improve discrimination against facial targets of different resolutions, The frame generation can be made suitable for different faces, and face detection trained by the face detection training method provided by the embodiment of the present invention can effectively improve the accuracy rate and enhance the robustness can do. The performance comparison between the face detection according to the embodiment of the present invention and the face detection trained by another method is as shown in Table 1 below.

  As can be seen from this, the embodiment of the present invention can improve the face detection discrimination ability of the face detection, and can improve the robustness of the face detection.

  The face detection training apparatus provided by the embodiment of the present invention will be described below, and the contents of the face detection training apparatus described later will be described in detail in order to implement the face detection training method provided by the embodiment of the present invention. It can be regarded as a program module necessary for the electronic device for performing the detection training, and the content of the face detection training device described later can be referred to in correspondence with the content of the face detection training method described above. .

FIG. 8 is a block diagram illustrating a configuration of a face detection training apparatus provided by an embodiment of the present invention. Referring to FIG.
Obtaining a batch data training sample of the current iteration, wherein the batch data training sample includes a plurality of training samples of different sample classes;
A sample center loss value determination module 200 for determining a center loss value corresponding to each training sample based on a feature vector of each training sample and a center feature vector of a sample class to which each training sample belongs;
A batch sample center loss value determination module 300 for determining a center loss value corresponding to the batch data training sample based on the center loss value corresponding to each of the training samples;
A detection target loss value determination module 400 for determining a target loss value for face detection based on at least the central loss value corresponding to the batch data training sample;
If the target loss value of the face detection does not reach the set training convergence condition, the network parameter of the face detection model is updated based on the target loss value of the face detection, and the parameter is updated to proceed to the next iteration. Module 500;
A detection output module 600 for outputting face detection when the target loss value for face detection reaches a set training convergence condition;
Can be included.

The detection target loss value determination module 400 determines a target loss value for face detection based on at least the central loss value corresponding to the batch data training sample.
Determine a target loss value for face detection based on the central loss value corresponding to the batch data training sample, the classification loss value corresponding to the batch data training sample, and the face frame coordinate regression loss value corresponding to the batch data training sample. May be included.

Further, the detection target loss value determination module 400 includes a central loss value corresponding to the batch data training sample, a classification loss value corresponding to the batch data training sample, and a face frame coordinate regression loss value corresponding to the batch data training sample. Based on the target loss value of face detection is determined, specifically,
A product of a central loss value corresponding to the batch data training sample and a first set weight, a product of a face frame coordinate regression loss value corresponding to the batch data training sample and a second set weight, and the batch data This may include summing the classification loss values corresponding to the training samples to obtain a target loss value for face detection.

The sample acquisition module 100 acquires a batch data training sample of the current iteration, and specifically,
Determining the target loss value corresponding to each training sample in the training sample set with the face detection model of the previous iteration;
Based on the target loss value of each training sample among the sample classes of the positive class in the training sample set, selecting the first number of training samples having the largest target loss value among the sample classes of the positive class, Selecting a second number of training samples having the largest target loss value among the sample classes of the negative class based on the target loss value of each training sample among the sample classes of the class, and selecting the first number of training samples for the second number; That the ratio value of the number corresponds to the set proportion,
The training sample selected from the sample class of the positive class and the training sample selected from the sample class of the negative class may constitute a batch data training sample of the current iteration.

In addition, the sample acquisition module 100 determines the target loss value corresponding to each training sample in the training sample set by the face detection model of the previous iteration, and specifically,
In the face detection model of the previous iteration, the central loss value, the classification loss value, and the face frame coordinate regression loss value corresponding to each training sample in the training sample set are obtained, among which the classification loss value corresponding to the training sample is Determined based on the classification prediction probability and the classification target probability corresponding to the training sample, and the face frame coordinate regression loss value corresponding to the training sample is determined based on the face frame prediction coordinate and the face frame target coordinate corresponding to the training sample. When,
Determining a target loss value for each training sample in the training sample set based on the central loss value, the classification loss value, and the face frame coordinate regression loss value corresponding to each training sample in the training sample set.

FIG. 9 shows another configuration of the face detection training apparatus provided by the embodiment of the present invention. Referring to FIGS. 8 and 9, the face detection training apparatus includes:
Determining a classification loss value corresponding to each training sample in the batch data training sample based on a classification prediction probability and a classification target probability corresponding to each training sample in the batch data training sample; May further include a batch sample classification loss value determination module 700 for determining a classification loss value corresponding to the batch data training sample based on the classification loss value corresponding to.

Further, the sample center loss value determination module 200 determines the center loss value corresponding to each training sample based on the feature vector of each training sample and the center feature vector of the sample class to which each training sample belongs.
Determining a feature vector of each training sample in the batch data training sample, and a central feature vector of each sample class in the batch data training sample;
For one training sample of the batch data training samples, determine a distance between a feature vector of the training sample and a center feature vector of a sample class to which the training sample belongs in the batch data training sample, and perform the training. Obtaining a center loss value corresponding to the sample.

Further, the sample center loss value determination module 200 determines a center feature vector of each sample class in the batch data training sample, and specifically,
For one sample class, determining each training sample belonging to the sample class in the batch data training sample;
Based on the feature vector of each training sample belonging to the sample class in the batch data training sample, determine the average value of the feature vector of each training sample belonging to the sample class, and determine the center of the sample class in the batch data training sample. Obtaining the update variable of the feature vector,
Obtaining a central feature vector of the sample class in the batch data training sample based on the update variable and the set learning rate.

FIG. 10 shows another configuration of the face detection training apparatus provided by the embodiment of the present invention. Referring to FIGS. 9 and 10, the face detection training apparatus comprises:
Based on the candidate frame regressor, the face frame predicted coordinates corresponding to each training sample in each of the batch data training samples are determined, the face frame predicted coordinates corresponding to each training sample, and the face frame target coordinates corresponding to each training sample. , A face frame coordinate regression loss value corresponding to each training sample is determined, and a face frame coordinate regression loss value corresponding to the batch data training sample is determined based on the face frame coordinate regression loss value corresponding to each training sample. The method may further include a batch sample face frame coordinate regression loss value determination module 800 for performing the process.

Further, the parameter updating module 500 updates the network parameters of the face detection model based on the target loss value of the face detection.
The method may include updating network parameters in the face detection model by back propagation based on the target loss value of face detection.

The parameter update module 500 updates the network parameters in the face detection model by back propagation based on the target loss value of the face detection.
Based on the target loss value of the face detection, and a network parameter in the face detection model of the previous iteration, to determine a parameter update value of the face detection,
Updating the network parameters in the face detection model of the previous iteration based on the parameter update value of the face detection.

Further, the face detection training apparatus provided by the embodiment of the present invention further includes:
A plurality of anchor frames covering different scales and aspect ratios are defined in advance, a sub-frame in a training sample is determined by the plurality of predefined anchor frames, and a candidate frame is predicted by the sub-frame.

Embodiments of the present invention further provide an electronic device, and a hardware configuration of the electronic device includes at least one memory and at least one processor, as shown in FIG.
A program is stored in the memory, and the processor calls the program, and according to the program,
Obtaining a batch data training sample of this iteration, wherein said batch data training sample includes a plurality of training samples of different sample classes;
Determining a central loss value corresponding to each training sample based on a feature vector of each training sample and a central feature vector of a sample class to which each training sample belongs;
Determining a central loss value corresponding to the batch data training sample based on the central loss value corresponding to each of the training samples;
Determining a target loss value for face detection based on at least the central loss value corresponding to the batch data training sample;
If the target loss value of the face detection has not reached the set training convergence condition, based on the target loss value of the face detection, update the network parameters of the face detection model, proceed to the next iteration,
When the target loss value of the face detection reaches the set training convergence condition, the training result of the face detection is output.

  Each example herein is described progressively, with each example primarily describing differences from the other embodiments, and the same or similar parts between each example may be referenced to each other. Since the device disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can refer to the description in the method section.

  One skilled in the art can further understand the following, wherein each exemplary unit and algorithm step described in connection with the embodiments disclosed herein may be implemented in electronic hardware, computer software, or To clarify hardware and software compatibility, which can be realized by a combination thereof, each exemplary component and step has been described generally in terms of functionality in the above description. . Whether these functions are performed in hardware or software depends on the particular application of the technical solution and the design constraints of the solution. Those skilled in the art will be able to use different ways to implement the described functionality for a particular application, but such implementation should not be considered beyond the scope of the present invention.

  The steps of a method or algorithm described in connection with the embodiments disclosed herein may be implemented directly in hardware, software modules executed by a processor, or a combination thereof. The software modules include a random access memory (RAM), a memory, a read only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, Or it can be located in any other form of storage medium known in the art.

  The above description of the disclosed embodiments allows those skilled in the art to make or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention should not be limited to these embodiments shown herein, but is in its broadest scope consistent with the principles and novel features disclosed herein.

100 Sample acquisition module 200 Sample center loss value determination module 300 Batch sample center loss value determination module 400 Detection target loss value determination module 500 Parameter update module 600 Detection output module 700 Batch sample classification loss value determination module 800 Batch sample face frame coordinate regression loss Value determination module

Claims (18)

A face detection training method,
Obtaining a batch data training sample of the iteration, wherein the batch data training sample includes a plurality of training samples of different sample classes;
Determining a central loss value corresponding to each training sample based on a feature vector of each training sample and a central feature vector of a sample class to which each training sample belongs;
Determining a central loss value corresponding to the batch data training sample based on the central loss value corresponding to each of the training samples;
Determining a target loss value for face detection based on the central loss value corresponding to the batch data training sample;
When the target loss value of the face detection reaches a set training convergence condition, outputting a face detection training result,
A face detection training method comprising: The method comprises:
If the target loss value of the face detection has not reached the set training convergence condition, based on the target loss value of the face detection, update the network parameters of the face detection model, and proceed to the next iteration,
The face detection training method according to claim 1, further comprising: Determining a target loss value for face detection based on the central loss value corresponding to the batch data training sample described above,
Determine a target loss value for face detection based on the central loss value corresponding to the batch data training sample, the classification loss value corresponding to the batch data training sample, and the face frame coordinate regression loss value corresponding to the batch data training sample. To do,
3. The face detection training method according to claim 1, further comprising: Based on the central loss value corresponding to the batch data training sample, the classification loss value corresponding to the batch data training sample, and the face frame coordinate regression loss value corresponding to the batch data training sample, a target loss value for face detection. To determine
A product of a central loss value corresponding to the batch data training sample and a first set weight, a product of a face frame coordinate regression loss value corresponding to the batch data training sample and a second set weight, and the batch data Summing the classification loss values corresponding to the training samples to obtain a target loss value for face detection;
The face detection training method according to claim 3, further comprising: The sample class of the plurality of training samples includes a positive class and a negative class, and obtaining the batch data training sample of the above-described iteration includes:
Determining the target loss value corresponding to each training sample in the training sample set in the model of the previous iteration;
Based on the target loss value of each training sample among the sample classes of the positive class in the training sample set, selecting the first number of training samples having the largest target loss value among the sample classes of the positive class, Selecting a second number of training samples having the largest target loss value among the sample classes of the negative class based on the target loss value of each training sample among the sample classes of the class, and selecting the first number of training samples for the second number; That the ratio of the numbers corresponds to the set proportion,
Constructing a batch data training sample for this iteration with the training sample selected from the positive class sample class and the training sample selected from the negative class sample class;
The face detection training method according to any one of claims 1 to 4, further comprising: Determining the target loss value corresponding to each training sample in the training sample set with the model of the previous iteration described above,
In the model of the previous iteration, a central loss value, a classification loss value, and a face frame coordinate regression loss value corresponding to each training sample in the training sample set are obtained, wherein the classification loss value corresponding to the training sample is the aforementioned The classification prediction probability and the classification target probability corresponding to the training sample are determined, and the face frame coordinate regression loss value corresponding to the training sample is determined based on the face frame prediction coordinate and the face frame target coordinate corresponding to the training sample. That
Determining a target loss value for each training sample in the training sample set based on the central loss value, the classification loss value, and the face frame coordinate regression loss value corresponding to each training sample in the training sample set;
The face detection training method according to claim 5, further comprising: The determination process of the classification loss value corresponding to the batch data training sample,
Determining a classification loss value corresponding to each training sample in the batch data training sample based on a classification prediction probability and a classification target probability corresponding to each training sample in the batch data training sample;
Determining a classification loss value corresponding to the batch data training sample based on a classification loss value corresponding to each training sample in the batch data training sample;
The face detection training method according to claim 3, further comprising: Confirmation processing of the face frame coordinate regression loss value corresponding to the batch data training sample,
Determining face frame prediction coordinates corresponding to each training sample in the batch data training sample;
Based on the face frame predicted coordinates corresponding to each training sample, and the face frame target coordinates corresponding to each training sample, determining a face frame coordinate regression loss value corresponding to each training sample;
Determining a face frame coordinate regression loss value corresponding to the batch data training sample based on the face frame coordinate regression loss value corresponding to each training sample;
The face detection training method according to claim 3, further comprising: Based on the feature vector of each training sample and the center feature vector of the sample class to which each training sample belongs, determining the central loss value corresponding to each training sample,
Determining a feature vector of each training sample in the batch data training sample, and a central feature vector of each sample class in the batch data training sample;
For a training sample in the batch data training sample, determine a distance between a feature vector of the training sample and a central feature vector of a sample class to which the training sample belongs in the batch data training sample, and correspond to the training sample. To obtain a central loss value
The face detection training method according to claim 1, comprising: Determining the central feature vector of each sample class in the batch data training sample described above,
For each sample class, determining each training sample belonging to the sample class in the batch data training sample;
Based on the feature vector of each training sample belonging to the sample class in the batch data training sample, determine the average value of the feature vector of each training sample belonging to the sample class, and determine the center of the sample class in the batch data training sample. Obtaining the update variable of the feature vector,
Obtaining a central feature vector of the sample class in the batch data training sample based on the update rate and the set learning rate;
10. The face detection training method according to claim 9, comprising: Updating network parameters in the model based on the target loss value of the face detection described above,
Updating network parameters in the face detection model by back propagation based on the target loss value of face detection;
The face detection training method according to any one of claims 2 to 4, comprising: Updating network parameters in the face detection model by back propagation based on the target loss value of the face detection described above,
Based on the target loss value of the face detection, and a network parameter in the face detection model of the previous iteration, to determine a parameter update value of the face detection,
Updating the network parameters in the face detection model of the previous iteration based on the parameter update value of the face detection;
The face detection training method according to claim 11, further comprising: The method comprises:
Pre-defining a plurality of anchor frames covering different scales and aspect ratios,
Determining a sub-frame in a training sample by the plurality of predefined anchor frames, and predicting a candidate frame by the sub-frame;
The face detection training method according to claim 1, further comprising: A sample acquisition module for acquiring batch data training samples comprising a plurality of training samples of different sample classes of the iteration;
A sample center loss value determination module for determining a center loss value corresponding to each training sample based on a feature vector of each training sample and a center feature vector of a sample class to which each training sample belongs;
A batch sample center loss value determination module for determining a center loss value corresponding to the batch data training sample based on the center loss value corresponding to each of the training samples;
A detection target loss value determination module for determining a target loss value for face detection based on the central loss value corresponding to the batch data training sample;
When the target loss value of the face detection reaches the set training convergence condition, a detection output module for outputting a training result of the face detection,
A face detection training apparatus comprising: The face detection training device,
If the target loss value of the face detection does not reach the set training convergence condition, the network parameter of the face detection model is updated based on the target loss value of the face detection, and the parameter is updated to proceed to the next iteration. The face detection training apparatus according to claim 14, further comprising a module. The sample acquisition module, specifically,
In the face detection model of the previous iteration, determine the target loss value corresponding to each training sample in the training sample set,
Based on the target loss value of each training sample among the sample classes of the positive class in the training sample set, selecting the first number of training samples having the largest target loss value among the sample classes of the positive class, Selecting a second number of training samples having the largest target loss value among the sample classes of the negative class based on the target loss value of each training sample among the sample classes of the class, and selecting the first number of training samples for the second number; The ratio of the numbers corresponds to the set proportion,
The training sample selected from the sample class of the positive class and the training sample selected from the sample class of the negative class constitute a batch data training sample of this iteration.
The face detection training apparatus according to claim 14, wherein: An electronic device including a memory and a processor,
A program is stored in the memory, and the processor calls the program, and according to the program,
Obtaining a batch data training sample of the iteration, wherein the batch data training sample includes a plurality of training samples of different sample classes;
Determining a central loss value corresponding to each training sample based on a feature vector of each training sample and a central feature vector of a sample class to which each training sample belongs;
Determining a central loss value corresponding to the batch data training sample based on the central loss value corresponding to each of the training samples;
Determining a target loss value for face detection based on the central loss value corresponding to the batch data training sample;
When the target loss value of the face detection reaches a set training convergence condition, a face detection training result is output,
Electronic equipment characterized by the above.   14. A computer readable storage medium containing instructions, wherein the instructions, when executed on a computer, cause the computer to perform the method according to any one of claims 1 to 13.